Hidden signals in online reviews: a machine learning framework for accurate sentiment influence

Purpose The purpose of this study is to enhance the accuracy and comprehensiveness of sentiment influence analysis in online reviews by addressing three critical limitations: improving sentiment classification accuracy, considering the combined effect of multiple emotions and mitigating biases in traditional usefulness voting methods. Design/methodology/approach This study employs a novel hybrid machine learning framework that integrates information entropy, principal component analysis (PCA) with K-means clustering and a bidirectional encoder representations from transformers-based sentiment classification model. Over 290,000 online reviews across four product categories were analyzed to predict perceived usefulness, calculate sentiment loss rates and systematically quantify sentiment influence. Findings The results demonstrate that different product types are influenced by positive, negative or mixed sentiments to varying extents. Specifically, economical products were predominantly influenced by positive sentiments, while applied products (e.g. medical apps) showed greater sensitivity to negative sentiments. Experiential and enjoyment products were equally influenced by both sentiments, highlighting the need for balanced strategies. Research limitations/implications The study’s data are primarily from the Chinese market; future research could extend the framework to international datasets. Additionally, although focused on e-commerce scenarios, the methodology could be expanded to other domains and integrated with hybrid models combining machine learning with regression testing. Practical implications Businesses can leverage this framework to optimize resource allocation effectively, target marketing strategies more accurately and improve customer satisfaction and loyalty by responding strategically to consumer emotions reflected in reviews. Social implications Accurate sentiment influence analysis empowers consumers with better information for decision-making and encourages businesses to align closely with genuine consumer needs and emotional feedback, thereby enhancing overall consumer welfare and market transparency. Originality/value This research pioneers the integration of qualitative sentiment classification with quantitative emotional influence measurement using machine learning. It introduces a novel metric – sentiment information loss rate – to quantify sentiment impacts precisely and explores the combined effect of multiple emotions, challenging the traditional one-to-one sentiment contagion model.